Types of AI development: from machine learning and LLMs to RAG and AI agents

“We need to add AI” sounds like a single technology decision. In practice, it could describe a forecasting model, a document-processing pipeline, an internal knowledge assistant, a tool-using agent or a system that coordinates several specialised agents.

Those approaches solve different problems and carry different costs, risks and operating requirements. A business that needs reliable invoice extraction should not automatically build an autonomous agent. A team that needs current answers from private documents should not expect a standalone large language model to know them.

The useful question is not which AI term is receiving the most attention. It is: what is the simplest type of AI development that can produce the required outcome dependably?

AI development is a spectrum, not a maturity ladder

Terms such as large language model, RAG, copilot, AI agent and agentic system are often presented as if each one replaces the previous technology. That is misleading. They describe different model capabilities, architectural patterns and levels of autonomy.

A production solution may combine several of them. For example, an internal support assistant could use a large language model to understand a question, retrieval-augmented generation to find approved knowledge, a workflow to check the response and a human approval step before anything is sent to a customer.

The main approaches can be understood as follows:

• Rules and workflow automation execute known, predefined logic.
• Predictive machine learning produces a score, forecast, recommendation or classification.
• Generative AI and LLM applications create, interpret or transform language and other content.
• Multimodal AI works across text, images, audio, video or documents.
• RAG supplies selected external information to a generative model when it answers.
• Assistants and copilots help a person complete work while leaving the person in control.
• AI workflows put model calls and tools into a designed sequence.
• AI agents dynamically choose tools and steps to pursue an outcome.
• Agentic and multi-agent systems coordinate longer-running, adaptive work with greater autonomy.

Fine-tuning is different again: it is a method for adapting a model’s behaviour, not a separate business solution.

Rules and workflow automation

Not every intelligent-looking process requires an AI model. If the rules are stable and the expected result is clear, conventional software automation is often faster, cheaper and easier to verify.

Examples include moving approved data between systems, sending a notification when a status changes, generating a scheduled report or routing a request according to known conditions.

AI becomes useful when a step contains ambiguity: interpreting an email, extracting inconsistent document content, matching a request to an appropriate category or drafting a response. The strongest systems frequently combine deterministic workflow controls with AI only where it adds value.

Predictive machine learning

Traditional machine learning learns patterns from historical data to predict or classify a future or unseen case. It remains the appropriate choice for many business problems even as generative AI receives more attention.

Common uses include:

• demand, revenue or workload forecasting;
• fraud, anomaly or equipment-failure detection;
• customer or enquiry classification;
• recommendations and prioritisation;
• risk and propensity scoring; and
• optimisation of schedules, stock or resources.

The critical inputs are suitable historical data, a measurable target and a process that can use the prediction. If the required output is a probability, score or category rather than newly generated content, predictive AI may be more appropriate than an LLM.

Generative AI and large language models

Large language models, or LLMs, are foundation models trained to interpret and generate language. An LLM application sends the model instructions and context, receives an output and places that capability inside a useful product or workflow.

They are well suited to tasks such as:

• summarising calls, documents or case histories;
• drafting emails, reports and knowledge articles;
• extracting structured information from unstructured text;
• classifying or routing enquiries;
• translating and transforming content; and
• answering questions from information supplied in the prompt.

A basic LLM does not automatically know a company’s current policies, customer records or private documents. It can also produce fluent statements that are unsupported or incorrect. Production development therefore needs clear instructions, appropriate context, output validation and evaluation against representative examples.

Multimodal AI, computer vision, speech and document processing

Modern AI systems are not limited to text. Multimodal models can work with combinations of text, images, audio, video and documents.

This broader category includes:

• Computer vision: inspecting, recognising or classifying visual information.
• Speech AI: transcribing calls, identifying speakers or generating spoken output.
• Document AI: combining OCR, layout understanding and language models to process forms, invoices and reports.
• Multimodal assistants: answering questions about images, diagrams, recordings or mixed document collections.

The model is only one component. Image quality, document layout, audio conditions, privacy requirements and the cost of human correction all affect whether the solution is commercially useful.

Retrieval-augmented generation

Retrieval-augmented generation, usually shortened to RAG, combines information retrieval with a generative model. When a person asks a question, the application searches an approved knowledge source and supplies the most relevant material to the LLM as context for its answer.

RAG is useful when responses must reflect current, private or specialist information, such as:

• internal policies and procedures;
• technical and product documentation;
• customer-support knowledge;
• contracts, reports or research collections; and
• information stored across business systems.

RAG is not a guarantee of truth. The answer can still be incomplete if the correct information is missing, inaccessible or poorly retrieved. Permissions should be applied before content reaches the model, and the system should show sources where users need to verify an answer.

Good RAG development includes data preparation, access control, search quality, document parsing, retrieval evaluation and monitoring—not only connecting an LLM to a vector database. Kitek’s data integration and reporting capability can also be relevant when the required information is scattered across operational systems.

Fine-tuning and custom models

Fine-tuning adapts a model using examples of the behaviour or output required. It can help when a task needs consistent terminology, formatting, classification or specialist response patterns that are difficult to achieve through instructions alone.

It is often confused with giving a model business knowledge. If information changes regularly, retrieval is usually more suitable because the source can be updated without retraining the model. Fine-tuning changes how a model behaves; RAG changes the information available when it responds.

Training a foundation model from scratch is a substantially different undertaking and is rarely the appropriate starting point for a small or mid-sized business. Existing models, combined with controlled data and workflow integration, can usually test the opportunity much sooner.

AI assistants, copilots and chatbots

An assistant or copilot helps a person complete a task but does not normally own the whole process. It may prepare a draft, surface relevant information, suggest a next step or turn a conversation into structured data.

Examples include:

• preparing a customer-response draft for approval;
• summarising an account before a service call;
• helping staff search policies and product knowledge;
• suggesting classifications or priorities; and
• assisting a user inside an existing business application.

This is frequently a sensible first production pattern. The AI reduces effort while a person remains responsible for the decision or external action. The system can collect feedback and evidence before greater automation is considered.

AI workflows

An AI workflow coordinates models, software rules and integrations through a predefined path. Unlike an open-ended agent, the major stages are designed in advance.

A document workflow might receive a file, extract its fields, check required values, send uncertain cases to a person and update the accounting system only after approval. A customer-enquiry workflow might classify the request, retrieve relevant information, draft a response and route it to the correct team.

Common patterns include:

• Chaining: the output of one step becomes the input to the next.
• Routing: requests are directed to different prompts, models or processes.
• Parallel review: several checks run independently before their results are combined.
• Evaluator and optimiser: one model produces an output while another checks it against defined criteria.

Workflows trade some flexibility for predictability. That is often the right decision when the process is known and auditability matters.

AI agents

An AI agent receives an objective, decides what steps to take, uses permitted tools, observes the results and continues until it reaches a stopping condition or needs human input.

An agent might search several information sources, query a business system, compare the results, create a recommendation and ask for approval before performing an action. The path can change depending on what the agent discovers.

Agents are most useful when the required steps cannot be fully known in advance. That flexibility also creates risk. Each additional decision or tool call can add latency, cost and another opportunity for error.

Production agents need:

• narrow, well-documented tools;
• least-privilege access to systems and data;
• limits on iterations, time and expenditure;
• human approval before consequential actions;
• complete logs and observable state;
• sandbox testing and realistic evaluations; and
• a safe way to stop, retry or escalate.

Agentic AI and multi-agent systems

Agentic AI is the broader term for systems in which models exercise some control over how work is completed. An agentic system may contain one agent, several agents, predefined workflows, memory, retrieval and access to external tools.

A multi-agent system assigns different roles to several agents and coordinates their work. One agent might plan, others might research specialised sources and another might evaluate the result.

This can be valuable when tasks genuinely require different capabilities or independent perspectives. It should not be the automatic target architecture. Multiple agents can duplicate work, pass errors between one another and make behaviour harder to reproduce or debug.

For many business applications, a single model call with good context—or a controlled workflow with retrieval and human review—will be more dependable than a highly autonomous multi-agent design.

How to choose the appropriate AI architecture

Start with the shape of the business problem:

• The logic is stable and predictable: use conventional automation.
• You need a forecast, score or category: evaluate predictive machine learning.
• You need to generate, interpret or transform content: start with an LLM or multimodal model.
• Answers must use current or private information: add retrieval and permission-aware grounding.
• A person should remain responsible: build an assistant or copilot.
• The task follows known stages: use an AI workflow.
• The route depends on intermediate results: consider a constrained agent.
• The task genuinely needs independent specialist roles: then consider multiple agents.

This is not a permanent choice. A business can begin with human-reviewed assistance, measure its performance and automate selected steps only when the evidence supports doing so.

What every production AI system still needs

Regardless of architecture, a dependable AI system needs normal software-engineering discipline plus controls for probabilistic behaviour.

That includes:

• representative test cases and measurable acceptance criteria;
• privacy, security and retention controls;
• authentication and permission-aware access;
• human escalation and review paths;
• monitoring of outputs, actions, cost and latency;
• failure, retry and rollback behaviour;
• versioning for prompts, models, tools and source data; and
• continuing evaluation as models and business information change.

A demonstration proves that a model can produce an impressive output. Production AI development must prove that the complete system behaves usefully and safely across ordinary cases, poor inputs and predictable failures.

Start with one bounded AI opportunity

The best first AI project is usually a defined workflow with a measurable result, available data and a clear owner. It should be narrow enough to test with representative cases and valuable enough that a successful result matters.

Begin with the workflow, define what success means and use the least complicated architecture that can meet it. Add retrieval, tools, autonomy or multiple agents only when each extra layer produces a demonstrable improvement.

Kitek helps businesses assess AI opportunities, prototype them safely and integrate useful capabilities into existing web, mobile and operational systems. Explore Kitek’s workflow automation and AI integration service →